Machine learning teams can improve their vision models with the help of a startup that was spun out of ETH in 2020. The company has raised a $12 million Series A funding round led by Atlantic Bridge and Open Ocean. btov Partners, which led the company's seed round last year, also took part in this round.
The company currently has more than 10 customers in both Europe and the U.S., including a number of large enterprises like the Swiss Federal Railways, and is currently running pilots with quite a few. The customer demand was the reason why LatticeFlow raised.
Tsankov met with some investors in Palo Alto while he was in the US. They noticed that we have a problem with customers who are new to us. You shouldn't run the company like we did, we had machine learning engineers support customers. Take $12 million and bring in some people. We saw that the market has changed when we spoke to other investors.
Most enterprises have a hard time bringing their models into production and then, when they do, they realize that they don't perform as well as they expected The promise of LatticeFlow is that it can identify potential blind spots. More than half a dozen critical blind spots were found in their state-of-the-art production models when it analyzed their data with a major medical company.
It is not enough to only look at the training data and ensure that there is a diverse set of images, but also look at the models.
Prof. Martin Vechev is a member of the founding team. The image is called Lattice Flow.
“If you only look at the data — and this is a fundamental differentiator for LatticeFlow because we not only find the standard data issues like labeling issues or poor-quality samples, but also model blind spots, which are the scenarios where the models are failing,” Tsankov explained. “Once the model is ready, we can take it, find various data model issues and help companies fix it.”
He said that models will often find hidden correlations that may confuse them and skew the results. The model would often label an image with a finger in it as a scratch if it was used to detect dents, scratches and other damage in images of cars. What's the reason? Customers in the training set would take a close-up picture with a scratch and point at it with their fingers. Even if there was no scratch on the car, the model would correlate fingers with scratches. There are issues that go beyond creating better labels and need a service that can look at both the model and the training data.
There is a bias in data for training models. A spurious feature is caused by the fact that people point at scratches. The issue was fixed with a custom augmentation. The image is called Lattice Flow.
It is not in the training business. Pre-trained models work with the service. It may offer a fully managed service in the future, too, as it uses the new funding to hire aggressively, both to better service its customers and to build out its product portfolio.
The operating partner at Atlantic Bridge said that most large-scale artificial intelligence models are not functioning reliably in the real world. The lack of tools that help engineers resolve critical artificial intelligence data and model errors is the main reason for this. This is the reason why the Atlantic Bridge team decided to invest in Lattice Flow. The company is poised for tremendous growth because it is the only company that can fix model defects at scale.
ETH spin-off LatticeFlow raises $2.8M to help build trustworthy AI systems
All Rights Reserved © 2024
